Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 13 additions & 6 deletions amber/src/main/python/core/models/internal_queue.py
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,6 @@
from threading import RLock
from typing import TypeVar, Set

from core.models.internal_marker import InternalMarker
from core.models.payload import DataPayload
from core.util.customized_queue.linked_blocking_multi_queue import (
LinkedBlockingMultiQueue,
Expand Down Expand Up @@ -77,11 +76,19 @@ def get(self) -> T:
def put(self, item: T) -> None:
if isinstance(item, InternalQueueElement):
if item.tag not in self._queue_ids:
self._queue.add_sub_queue(item.tag, 1 if item.tag.is_control else 2)
self._queue_ids.add(item.tag)
if isinstance(item, (DataElement, InternalMarker, ECMElement)):
self._queue.put(item.tag, item)
elif isinstance(item, DCMElement):
# registration must not interleave with disable_data/enable_data
with self._lock:
Comment on lines +79 to +80

@Yicong-Huang Yicong-Huang Jul 22, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

can we avoid locking? the whole design is to use a lock free data structure in the hot path (e.g., put and get)

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think the lock might not quite be on the hot path as it's only reachable on a channel's first put. I instrumented it to verify: for 100k puts across 10 channels acquire the lock exactly 10 times. Also, I think the layer below already locks on every element: SubQueue.put takes put_lock and get takes take_lock on each call. I did try a lock-free version (publish the channel id first, then re-check _queue_state), but there's a window where the element lands in the still-enabled sub-queue before disable_data gets to it, which reopens the same leak this PR fixes. The lock-free issues could probably be worked around by creating the sub-queue disabled and only enabling it after rechecking the state, switching _queue_ids to copy-on-write, but that version might be overcomplicated. This lock might feel like the simpler safe choice. Happy to discuss!

if item.tag not in self._queue_ids:
self._queue.add_sub_queue(
item.tag, 1 if item.tag.is_control else 2
)
# while data is disabled, a new data sub-queue must
# start disabled too (before its first element is
# enqueued), or it would leak data during the pause
if not item.tag.is_control and self._queue_state:
self._queue.disable(item.tag)
self._queue_ids.add(item.tag)
if isinstance(item, (DataElement, ECMElement, DCMElement)):
self._queue.put(item.tag, item)
else:
raise ValueError(f"item {item} is not recognized by internal queue")
Expand Down
195 changes: 195 additions & 0 deletions amber/src/test/python/core/models/test_internal_queue.py
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@
# specific language governing permissions and limitations
# under the License.

import threading
from dataclasses import dataclass

import pytest
Expand Down Expand Up @@ -364,3 +365,197 @@ def test_it_can_disable_and_enable_a_single_data_channel(
queue.enable(data_channel)
assert queue.get() is blocked
assert queue.is_empty()

# Regression tests below: data channels whose sub-queue is created lazily
# (on the channel's first put) AFTER disable_data has been called must
# come up disabled — a paused or backpressured worker must not be able to
# dequeue data from them, and is_data_enabled() must not flip back to
# True just because a new channel delivered its first message.

def test_channel_registered_after_disable_comes_up_disabled(
self, queue, data_channel
):
# the main regression: disable first, then the channel's FIRST put
queue.disable_data(InternalQueue.DisableType.DISABLE_BY_PAUSE)
queue.put(self.data_element(data_channel))
assert not queue.is_data_enabled()
# the element stays queued but must not be dequeuable
assert queue.size_data() == 1
assert queue._queue.peek() is None

@pytest.mark.timeout(2)
def test_enable_data_releases_a_channel_registered_mid_disable(
self, queue, data_channel
):
queue.disable_data(InternalQueue.DisableType.DISABLE_BY_PAUSE)
data = self.data_element(data_channel)
queue.put(data)
assert queue.enable_data(InternalQueue.DisableType.DISABLE_BY_PAUSE)
assert queue.is_data_enabled()
assert queue._queue.peek() is data
assert queue.get() is data
assert queue.is_empty()

@pytest.mark.timeout(2)
def test_channel_registered_under_stacked_disables_stays_disabled(
self, queue, data_channel
):
queue.disable_data(InternalQueue.DisableType.DISABLE_BY_PAUSE)
queue.disable_data(InternalQueue.DisableType.DISABLE_BY_BACKPRESSURE)
data = self.data_element(data_channel)
queue.put(data)
# releasing only one of the two reasons must not open the channel
assert not queue.enable_data(InternalQueue.DisableType.DISABLE_BY_PAUSE)
assert not queue.is_data_enabled()
assert queue._queue.peek() is None
# releasing the remaining reason makes the element dequeuable
assert queue.enable_data(InternalQueue.DisableType.DISABLE_BY_BACKPRESSURE)
assert queue.is_data_enabled()
assert queue.get() is data

@pytest.mark.timeout(2)
def test_control_channel_registered_mid_disable_is_never_blocked(
self, queue, control_channel, data_channel
):
# register a data channel first so is_data_enabled() is meaningful
queue.put(self.data_element(data_channel))
queue.disable_data(InternalQueue.DisableType.DISABLE_BY_PAUSE)
# the control channel's FIRST put happens while data is disabled
dcm = self.dcm_element(control_channel)
queue.put(dcm)
# control must flow immediately, and data must stay disabled
assert queue._queue.peek() is dcm
assert queue.get() is dcm
assert not queue.is_data_enabled()
assert queue.size_data() == 1

@pytest.mark.timeout(2)
def test_channel_registered_before_disable_is_disabled_and_reenabled(
self, queue, data_channel
):
# baseline: the pre-existing behavior for eagerly-registered channels
data = self.data_element(data_channel)
queue.put(data)
queue.disable_data(InternalQueue.DisableType.DISABLE_BY_PAUSE)
assert not queue.is_data_enabled()
assert queue._queue.peek() is None
assert queue.enable_data(InternalQueue.DisableType.DISABLE_BY_PAUSE)
assert queue.is_data_enabled()
assert queue.get() is data

@pytest.mark.timeout(2)
def test_channel_registered_while_enabled_behaves_normally(
self, queue, second_data_channel
):
data = self.data_element(second_data_channel)
queue.put(data)
assert queue.is_data_enabled()
assert queue._queue.peek() is data
assert queue.get() is data
assert queue.is_empty()

@pytest.mark.timeout(10)
def test_concurrent_first_time_puts_while_toggling_disable(self, queue):
# concurrency smoke test: receiver threads register brand-new data
# channels while the DP thread toggles disable_data/enable_data;
# only the final state is asserted, deterministically.
n_threads = 8
elements_per_thread = 25
start_barrier = threading.Barrier(n_threads + 1)
errors = []

def producer(thread_index):
channel = ChannelIdentity(
ActorVirtualIdentity(f"upstream_{thread_index}"),
ActorVirtualIdentity("dummy_worker_id"),
False,
)
try:
start_barrier.wait()
for _ in range(elements_per_thread):
queue.put(self.data_element(channel))
except Exception as exc: # pragma: no cover - failure path
errors.append(exc)

threads = [
threading.Thread(target=producer, args=(i,)) for i in range(n_threads)
]
for thread in threads:
thread.start()
start_barrier.wait()
for _ in range(5):
queue.disable_data(InternalQueue.DisableType.DISABLE_BY_PAUSE)
queue.enable_data(InternalQueue.DisableType.DISABLE_BY_PAUSE)
for thread in threads:
thread.join()
# one last full cycle after all puts settled: every channel must be
# disabled, then re-enabled with its count added back exactly once
queue.disable_data(InternalQueue.DisableType.DISABLE_BY_PAUSE)
assert not queue.is_data_enabled()
assert queue._queue.peek() is None
assert queue.enable_data(InternalQueue.DisableType.DISABLE_BY_PAUSE)

assert not errors
total = n_threads * elements_per_thread
assert queue.is_data_enabled()
assert queue.size_data() == total
# size() is the getable total_count: a mismatch with size_data()
# means an element was double-counted or lost by a toggle race
assert queue.size() == total
drained = 0
while queue._queue.peek() is not None:
queue.get()
drained += 1
assert drained == total
assert queue.is_empty()

@pytest.mark.timeout(30)
def test_concurrent_first_time_puts_racing_disable_enable_toggles(self):
# Receiver threads deliver first-ever messages on distinct new data
# channels while the DP-thread side toggles pause on and off. Only
# the final state is asserted (deterministic): with the queue left
# disabled, nothing is dequeuable; after the final enable_data every
# element is dequeuable exactly once, so total_count stayed exact.
threads, channels_per_thread, toggles = 4, 10, 10
for _ in range(5):
queue = InternalQueue()
errors = []
start = threading.Barrier(threads + 1)

def producer(thread_id):
try:
start.wait()
for i in range(channels_per_thread):
channel = ChannelIdentity(
ActorVirtualIdentity(f"upstream-{thread_id}-{i}"),
ActorVirtualIdentity("dummy_worker_id"),
False,
)
queue.put(self.data_element(channel))
except Exception as exc: # pragma: no cover - failure path
errors.append(exc)

producers = [
threading.Thread(target=producer, args=(t,)) for t in range(threads)
]
for producer_thread in producers:
producer_thread.start()
start.wait()
for _ in range(toggles):
queue.disable_data(InternalQueue.DisableType.DISABLE_BY_PAUSE)
queue.enable_data(InternalQueue.DisableType.DISABLE_BY_PAUSE)
queue.disable_data(InternalQueue.DisableType.DISABLE_BY_PAUSE)
for producer_thread in producers:
producer_thread.join()

assert errors == []
total = threads * channels_per_thread
assert queue.size_data() == total
assert queue._queue.peek() is None
assert not queue.is_data_enabled()
assert queue.enable_data(InternalQueue.DisableType.DISABLE_BY_PAUSE)
dequeued = 0
while queue._queue.peek() is not None:
queue.get()
dequeued += 1
assert dequeued == total
Loading